Code Mode v2 — rollout tracker

[pratikbin]

Tracker for the v2 cutover. PR: #11.

Cross-repo dependency

Backend gate must merge first: autogen-backend-v2#451 (private) — adds operationId CI gate required by the MCP sidecar's typed SDK codegen.

Pre-merge

• Backend #451 merged
• MCP feat(mcp)!: v2.0.0 — Code Mode (workerd sidecar, 10-tool surface) #11 reviews complete
• CI green on MCP feat(mcp)!: v2.0.0 — Code Mode (workerd sidecar, 10-tool surface) #11 (token-budget gate, existing workflows)
• MCP feat(mcp)!: v2.0.0 — Code Mode (workerd sidecar, 10-tool surface) #11 marked Ready for Review (un-Draft)

Build + publish images

• docker build -t <registry>/createos-mcp:v2.0.0-rc1 .
• docker build -t <registry>/createos-mcp-workerd:v2.0.0-rc1 codemode/workerd
• Push both to internal registry
• (Optional) Mirror createos-mcp:v2.0.0-rc1 to a public mirror if community needs it

Staging

• Apply k8s/pod-v2.yaml in staging
• Pod becomes Ready (/readyz 200, workerd /health 200)
• bun scripts/smoke.ts against staging URL: ok=50/50, p50<500 ms, p95<2000 ms
• Manual: real Claude Code session against staging-v2 — deploy a sample Next.js project with ≤6 MCP tool calls
• bun scripts/soak.ts — failed=0 over 1000 requests
• Internal team usage for 1 week. Track:
  • tool-call distribution (search vs execute vs pollJob vs natives)
  • ApiError rate
  • wall-timeout count (should be 0 for typical workloads)
  • hot endpoints among the 87 long-tail (consider promoting back to native if any clearly belong on the fast path)

Pre-prod

• No regressions vs v1 over the 1-week soak
• Tag prior v1 commit on main as v1.x.y-final
• Build + push final createos-mcp:v2.0.0 + createos-mcp-workerd:v2.0.0 images

Production rollout (blue-green)

• 5% traffic → wait 30 min, watch metrics
• 25% → wait 1 hour
• 50% → wait 2 hours
• 100%
• Halt + rollback if errorKind:infra or overall error rate rises above v1 baseline at any step

Post-rollout

• Decommission v1 image after 1 incident-free week
• Tag v2.0.0 on main
• Merge feat/code-mode-v2 → main
• Update README to reflect v2 as the supported surface

Known limitations

• workerd 2026 has no synchronous isolate.terminate(). Wall-timeout (600 s) and dispatcher dispose sever the api-proxy callback, so runaway code's further calls surface as ProxyError, but the V8 isolate continues until the Worker Loader's LRU evicts it. Upstream tracking: cloudflare/workerd.
• ConfigMap for MCP config (mounted at /etc/createos) must be created in each cluster; example structure documented in k8s/pod-v2.yaml.

Linked

• MCP PR: feat(mcp)!: v2.0.0 — Code Mode (workerd sidecar, 10-tool surface) #11
• Backend PR: NodeOps-app/autogen-backend-v2#451 (private)
• Backend ops tracker: NodeOps-app/autogen-backend-v2#452 (private)